P4 to Git Change 1456988 by gandryey@gera-w8 on 2017/09/08 17:48:25

SWDEV-129129 - [[CQE OCL][Vega vs Fiji] Upto 12% Performance drop observed on VEGA10 compared to FIJI while running BlackMagic Davinci Resolve
	- Force tiny read_only buffers into USWC memory. That will avoid expensive tiny data uploads, which occur every frame.

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/pal/paldevice.cpp#59 edit
... //depot/stg/opencl/drivers/opencl/runtime/utils/flags.hpp#278 edit


[ROCm/clr commit: caa25fc792]

projects/clr/rocclr/runtime/device/pal/paldevice.cpp

@@ -1228,6 +1228,12 @@ pal::Memory* Device::createBuffer(amd::Memory& owner, bool directAccess) const {
       ? Resource::Remote
       : Resource::Local;
 
+  // Check if runtime can force a tiny buffer into USWC memory
+  if ((size <= (GPU_MAX_REMOTE_MEM_SIZE * Ki)) && (type == Resource::Local) &&
+      (owner.getMemFlags() & CL_MEM_READ_ONLY)) {
+    type = Resource::RemoteUSWC;
+  }
+
   if (owner.getMemFlags() & CL_MEM_BUS_ADDRESSABLE_AMD) {
     type = Resource::BusAddressable;
   } else if (owner.getMemFlags() & CL_MEM_EXTERNAL_PHYSICAL_AMD) {

projects/clr/rocclr/runtime/utils/flags.hpp

@@ -187,6 +187,8 @@ release(bool, OCL_STUB_PROGRAMS, false,                                       \
         "1 = Enables OCL programs stubing")                                   \
 release(bool, GPU_ANALYZE_HANG, false,                                        \
         "1 = Enables GPU hang analysis")                                      \
+release(uint, GPU_MAX_REMOTE_MEM_SIZE, 2,                                     \
+        "Maximum size (in Ki) that allows device memory substitution with system") \
 release_on_stg(uint, GPU_WAVE_LIMIT_CU_PER_SH, 0,                             \
         "Assume the number of CU per SH for wave limiter")                    \
 release_on_stg(uint, GPU_WAVE_LIMIT_MAX_WAVE, 10,                             \